Agentic AI tools, represented by OpenClaw, are driving memory market demand into a new paradigm. According to market intelligence, Morgan Stanley's latest report on March 18 indicates that as AI transitions from "thinking" to "executing," DRAM will replace HBM as the most challenging chip bottleneck in AI infrastructure, extending the memory cycle far beyond expectations.
Channel surveys suggest server DRAM DDR5 prices are projected to increase by over 50% quarter-over-quarter in Q2 2026, with some Chinese hyperscale cloud providers offering even higher bids. DDR4 contract prices are expected to rise 40%-50%, while enterprise SSD pricing for NAND is forecast to climb no less than 40%-50%. Morgan Stanley believes the memory industry is currently in the middle of an upcycle, with supply constraints tighter than previously anticipated—"Wall Street's earnings forecasts will have to catch up with reality."
This outlook is directly reflected in target price adjustments: SK Hynix's 2026-2027 EPS forecasts were raised by 24% and 32% respectively, with its target price increased from 1.1 million won to 1.3 million won, implying 43% upside from current levels. Samsung Electronics' common stock target was raised to 251,000 won, with both stocks maintaining "Overweight" ratings.
Morgan Stanley's core thesis is that markets are accustomed to linear thinking, while AI capabilities are expanding at an exponential pace. As AI evolves from "generating answers" to "completing tasks," the magnitude of memory demand will leap accordingly, and this transformation is only beginning to accelerate.
"Executing tasks consumes more DRAM than thinking," Morgan Stanley's report notes as its foundational premise. Traditional large language models operate via GPU-dominated linear processes: receiving queries, processing input tokens (prefill phase), then generating responses token-by-token (decode phase), with CPUs handling text conversion. In this workflow, GPU computing power is the decisive bottleneck, while DRAM merely assists with cache operations.
Agentic AI fundamentally alters this dynamic. OpenClaw, for example, is an open-source, self-hosted AI assistant that can simultaneously connect to over 50 messaging platforms including WhatsApp and Telegram, with system-level permissions for browser automation, file operations, command execution, and API calls. It doesn't just "answer questions" but "completes tasks"—searching the web, reading documents, calling external tools, and executing code to deliver multi-step collaborative results.
This paradigm shift means workflows expand from single GPU inferences to multi-step coordination, tool calls, and orchestration, where CPU computation time often contributes more to overall latency than GPU processing. Meanwhile, multiple agents must continuously share context, offload KV caches, and store/retrieve intermediate results—elevating memory from a supporting role to the core bottleneck.
Morgan Stanley's detailed analysis of OpenClaw's memory requirements concludes that DRAM dominates all other hardware constraints in such agentic tools. The tool operates in two distinct modes:
Lightweight gateway mode (remotely calling APIs like Claude or GPT-4): Even here, the bottleneck shifts from GPU/CPU to Node.js runtime DRAM usage. Minimum requirement is 2GB DRAM, with 4GB recommended for stable production use.
Local model mode (running AI models directly on-device): DRAM and graphics HBM become dual constraints. Morgan Stanley recommends 32GB system DRAM; running 7B-8B parameter models requires additional 8GB graphics DRAM, 13B-70B models need 16-24GB, while超大 models like Llama 3 70B demand over 80GB.
The report emphasizes that memory insufficiency doesn't cause performance degradation but immediate failure—JavaScript throws "heap out of memory" errors, leading to installation failures and operational crashes. This highlights memory's hard constraint nature in agentic scenarios: insufficient memory means system death, not slowdown.
OpenClaw's memory demand pattern reflects a broader structural shift. Morgan Stanley observes AI computing bottlenecks systematically migrating from raw computation to data movement, from HBM to system DRAM, with memory architecture evolving from HBM-centric to multi-tiered HBM-DRAM-NVMe SSD combinations.
One technical driver is rapidly expanding long-context needs. KV caches grow linearly with token counts and require network transmission in disaggregated inference scenarios, significantly increasing CPU I/O burdens. Core agentic operations like RAG retrieval and context management involve intensive memory I/O.
Market confirmation is equally clear. Intel and AMD recently acknowledged substantial shortages of high-core-count server processors. AMD EPYC CPU revenue exceeded 40% of total server CPU sales for the first time, with EPYC-based cloud instance deployments growing over 50% YoY. NVIDIA launched separately sold Vera CPUs and signed a multi-year Meta agreement to deploy standalone CPUs for personal agent operations at scale.
These structural shifts are materially impacting pricing. For DRAM, limited spot transactions for server DDR5 in Q2 2026 already show 50% quarterly increases, accepted by hyperscalers with even higher bids from Chinese providers. By late February, 64GB RDIMM contract prices reached $910-920, about 20% above Q1's $800 average. LPDDR and consumer DRAM prices are projected to rise at least 40%-50% in Q2, while HBM3E—previously expected to decline 20%-25%—is now seeing mid-single-digit percentage increases in ASIC customer renewals.
For NAND, enterprise SSD pricing is forecast to rise 40%-50% quarterly in Q2, with consumer products increasing no less than 60%. In some scenarios, eSSD prices could double again in Q2.
Morgan Stanley maintains that YoY price acceleration continues, with the industry still mid-cycle. Once market earnings forecasts adjust to reflect unprecedented capacity constraints, significant valuation repair potential exists for related stocks, while potential capital return increases could further support outperformance.
Comments